Unsupervised Anomaly Detection for IoT Time Series Signals with GANs

The Internet of Things (IoT) enables interconnection of heterogeneous devices through wireless and mobile communication technologies, with sensing devices continuously generating a large amount of time series signals. Extracting valuable information from this data has become increasingly challenging. In this context, anomaly detection techniques have emerged to identify events of interest. This paper proposes UatGAN, an unsupervised anomaly detection method for IoT time series signals utilizing generative adversarial networks (GANs). The method combines autoencoders (AEs) and GANs, using the encoderdecoder structure of AE to learn compressed representations of input data, and enhances sensitivity to anomalous inputs through adversarial training of GANs. Considering time correlation, the dynamic time warping (DTW) algorithm is introduced to calculate reconstruction errors, and a new anomaly diagnosis strategy is proposed. Experiments conducted on public datasets demonstrate that our method detects anomalies more accurately than baseline methods.